src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp

/*
 * Copyright (c) 2019-2020, 2022-2024 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#ifdef __aarch64__

#include "arm_gemm.hpp"

#include "kernels/a64_gemm_s16_8x12.hpp"
#include "kernels/a64_gemm_s8_4x4.hpp"
#include "kernels/a64_gemm_s8_8x12.hpp"
#include "kernels/a64_hybrid_s8qa_dot_4x16.hpp"
#include "kernels/a64_hybrid_s8qa_mmla_4x16.hpp"
#include "kernels/a64_hybrid_s8qs_dot_6x16.hpp"
#include "kernels/a64_hybrid_s8qs_mmla_6x16.hpp"
#include "kernels/a64_hybrid_s8s32_dot_6x16.hpp"
#include "kernels/a64_hybrid_s8s32_mmla_6x16.hpp"
#include "kernels/a64_interleaved_s8s32_mmla_8x12.hpp"
#include "kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp"
#include "kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp"

#ifdef ARM_COMPUTE_ENABLE_SVE
#ifdef ARM_COMPUTE_ENABLE_SME2
#include "kernels/sme2_gemv_s8qa_dot_16VL.hpp"
#include "kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp"
#include "kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp"
#include "kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp"
#endif // ARM_COMPUTE_ENABLE_SME2

#include "kernels/sve_hybrid_s8qa_dot_4x4VL.hpp"
#include "kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp"
#include "kernels/sve_hybrid_s8qs_dot_6x4VL.hpp"
#include "kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp"
#include "kernels/sve_hybrid_s8s32_dot_6x4VL.hpp"
#include "kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp"
#include "kernels/sve_interleaved_s8s32_dot_8x3VL.hpp"
#include "kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp"
#endif // ARM_COMPUTE_ENABLE_SVE

#include "gemm_hybrid_indirect.hpp"
#include "gemm_hybrid_quantized.hpp"
#include "gemm_hybrid_quantized_inline.hpp"
#include "gemm_interleaved.hpp"
#include "gemv_pretransposed.hpp"
#include "quantize_wrapper.hpp"
#include "utils.hpp"

namespace arm_gemm {

static const GemmImplementation<int8_t, int8_t, Requantize32> gemm_qint8_methods[] =
{
#ifdef ARM_COMPUTE_ENABLE_SVE
#ifdef ARM_COMPUTE_ENABLE_SME2
{
    GemmMethod::GEMM_HYBRID,
    "sme2_gemv_s8qa_dot_16VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && quant_hybrid_asymmetric(qp) && args._Msize == 1 && !args._indirect_input && args._nbatches == 1;  },
    nullptr,
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemvPretransposed<cls_sme2_gemv_s8qa_dot_16VL, int8_t, int8_t, Requantize32>(args, qp); }
},
{
    GemmMethod::GEMM_INTERLEAVED,
    "sme2_interleaved_nomerge_s8q_mopa_1VLx4VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
    [](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<int32_t>();
                               return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_1VLx4VL, int8_t, int8_t>(args, qp); }
},
{
    GemmMethod::GEMM_INTERLEAVED,
    "sme2_interleaved_nomerge_s8q_mopa_4VLx1VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
    [](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<int32_t>();
                               return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL, int8_t, int8_t>(args, qp); }
},
{
    GemmMethod::GEMM_INTERLEAVED,
    "sme2_interleaved_nomerge_s8q_mopa_2VLx2VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
    nullptr,
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL, int8_t, int8_t>(args, qp); }
},
#endif // ARM_COMPUTE_ENABLE_SME2
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "sve_hybrid_s8qa_mmla_4x4VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "sve_hybrid_s8qs_mmla_6x4VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_symmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "sve_interleaved_s8s32_mmla_8x3VL",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "sve_hybrid_s8s32_mmla_6x4VL",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "sve_hybrid_s8qs_dot_6x4VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_symmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "sve_hybrid_s8qa_dot_4x4VL",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "sve_hybrid_s8s32_dot_6x4VL",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "sve_interleaved_s8s32_dot_8x3VL",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && (args._Ksize>4); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>(args, qp); }
),
#endif // ARM_COMPUTE_ENABLE_SVE
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "a64_hybrid_s8qa_mmla_4x16",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "a64_hybrid_s8qs_mmla_6x16",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_symmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "a64_interleaved_s8s32_mmla_8x12",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "a64_hybrid_s8s32_mmla_6x16",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>(args, qp); }
),
{
    GemmMethod::GEMM_HYBRID_QUANTIZED,
    "a64_smallK_hybrid_s8s32_dot_8x4",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int8_t>(args, qp); }
},
{
    GemmMethod::GEMM_HYBRID_QUANTIZED,
    "a64_smallK_hybrid_s8s32_dot_6x4",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int8_t>(args, qp); }
},
{
    GemmMethod::GEMM_INTERLEAVED,
    "a64_gemm_s16_8x12",
    nullptr,
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s16_8x12, int8_t, int8_t>(args, qp); }
},
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "a64_hybrid_s8qs_dot_6x16",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_symmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "a64_hybrid_s8qa_dot_4x16",
    [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_asymmetric(qp); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_HYBRID,
    "a64_hybrid_s8s32_dot_6x16",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "a64_gemm_s8_8x12",
    [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); },
    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>(args, qp); }
),
GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
    GemmMethod::GEMM_INTERLEAVED,
    "a64_gemm_s8_4x4",
    nullptr,
    [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>::estimate_cycles<int8_t>(args); },
    [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>(args, qp); }
),
{
    GemmMethod::QUANTIZE_WRAPPER,
    "quantized_wrapper",
    [](const GemmArgs &args, const Requantize32 &) { return !args._indirect_input; },
    [](const GemmArgs &, const Requantize32 &) { return false; },
    [](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<int8_t, int8_t, int32_t>(args, qp); }
},
{
    GemmMethod::DEFAULT,
    "",
    nullptr,
    nullptr,
    nullptr
}
};

template<>
const GemmImplementation<int8_t, int8_t, Requantize32> *gemm_implementation_list<int8_t, int8_t, Requantize32>() {
    return gemm_qint8_methods;
}

template UniqueGemmCommon<int8_t, int8_t> gemm<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
template bool has_opt_gemm<int8_t, int8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os);
template KernelDescription get_gemm_method<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);

} // namespace arm_gemm

#endif // __aarch64__